Published in Marketing Science, Articles in Advance 22 Feb 2016 This is an Author Accepted Manuscript (AAM) pre-print. For the version of record, see DOI http://dx.doi.org/10.1287/mksc.2015.0968 Mining Brand Perceptions from Twitter Social Networks Aron Culotta Department of Computer Science, Illinois Institute of Technology, Chicago, IL, 60616 [email protected] Jennifer Cutler Kellogg School of Management, Northwestern University, Evanston, IL, 60208 [email protected] Consumer perceptions are important components of brand equity and therefore marketing strategy. Seg- menting these perceptions into attributes such as eco-friendliness, nutrition, and luxury enable a fine-grained understanding of the brand's strengths and weakness. Traditional approaches towards monitoring such per- ceptions (e.g., surveys) are costly and time-consuming, and their results may quickly become outdated. Extant data mining methods are not suitable for this goal, and generally require extensive hand-annotated data or context customization, which leads to many of the same limitations as direct elicitation. Here, we investigate a novel, general, and fully automated method for inferring attribute-specific brand perception ratings by mining the brand's social connections on Twitter. Using a set of over 200 brands and three per- ceptual attributes, we compare the method's automatic ratings estimates with directly-elicited survey data, finding a consistently strong correlation. The approach provides a reliable, flexible, and scalable method for monitoring brand perceptions, and offers a foundation for future advances in understanding brand-consumer social media relationships. Key words : social media; brand image; market structure; Twitter; attribute ratings; perceptual maps; big data; data mining; social networks 1. Introduction Understanding how consumers perceive brands is fundamental to much of marketing strat- egy. A central analytical tool used to do so is perceptual mapping, which organizes brands according to how consumers rate them with respect to attributes such as eco-friendliness or luxury (Green et al. 1989, Shocker and Srinivasan 1979, Steenkamp et al. 1994). Con- sumer ratings are typically collected through surveys or other elicitation means (Aaker 1996, Hauser and Koppelman 1979, Lehmann et al. 2008, Steenkamp and Van Trijp 1997); however, these data are costly and time-consuming to collect and may quickly become outdated. 1 Culotta and Cutler: Mining Brand Perceptions from Twitter Social Networks 2 Article submitted to Marketing Science; manuscript no. (Please, provide the manuscript number!) The recent proliferation of social media use by both marketers and consumers offers a promising data source to understand consumer perceptions; yet the noise, volume, and ambiguity of such data pose substantial challenges to algorithmic solutions. In this paper, we introduce and validate a fully-automated and highly generalizable method for esti- mating brand perceptions along a perceptual attribute1 of choice from publicly available secondary social media data, specifically Twitter. To the best of our knowledge, there are no extant data mining approaches developed for this task. At a high level, our algorithm takes as input a brand name and a query specifying the attribute of interest (e.g., \eco-friendliness"). It then returns a real value indicating the strength of association between the brand and the attribute. The main source of evidence used by our approach is the similarity between a brand's Twitter account and a set of exemplar accounts representing a perceptual attribute | e.g., the similarity between Smart Automobile's account and those of the EPA and GreenPeace may signal its perceived eco-friendliness. The method we develop is innovative in several ways. First, while most extant methods analyze user-generated text, we instead rely only on the structure of the brand's social network, which offers advantages in simplicity and scale. Second, we focus our analysis on the platform Twitter, which has received limited attention in the marketing literature, but offers advantages in data relevance and accessibility. Third, we introduce a fully-automated and highly generalizable process that requires only a keyword as input to generate near real-time estimates of brand ratings for an attribute mapping to that keyword. By leveraging the crowd-organization of social media, we circumvent the often extensive manual tuning and customization requirements of extant data mining approaches, thus providing a versatile and scalable method that can be applied to a range of marketing inquiries. To validate the effectiveness of the method, we use it to estimate perceptual ratings along three example attributes (eco-friendliness, luxury, and nutrition) for over two hundred brands across four sectors2 (Apparel, Cars, Food & Beverage, and Personal Care), and collect directly elicited survey ratings for the same set of brands and attributes. We find 1 For consistency, we use the word attribute throughout the paper, though we mean it to include any specific aspect of brand identity that can be identified through a keyword and rated along a continuum. These might also map to perceptual dimensions or associations, as they are referred to in other areas of the literature. 2 As will be explained in x4, attributes were tested only for sectors that made sense{e.g., nutrition perceptions were not estimated for car brands. Culotta and Cutler: Mining Brand Perceptions from Twitter Social Networks Article submitted to Marketing Science; manuscript no. (Please, provide the manuscript number!) 3 an average correlation over all sector-attribute combinations of 0.72, indicating that this fully-automated approach provides a reliable signal of current brand perceptions. This correlation meets or exceeds standards set in prior literature for related (though distinct) tasks, despite the manual customization required to implement these extant methods. Our core contribution, then, is a new methodological tool to quantify consumer percep- tions of brands with respect to a specified attribute. Our approach is a real-time, low-cost alternative to extant methods that firms and researchers can use for a number of com- mon marketing tasks, such as generating perceptual maps, monitoring market structures, and informing research models (Green 1975, Green et al. 1989, Hauser and Simmie 1981, Schmalensee and Thisse 1988). In addition to the algorithm itself, our scientific contribu- tion consists of a multi-faceted empirical validation against primary survey data, including an exploration of how a number of algorithmic variants affect accuracy. In the next section, we discuss relevant work from the marketing literature, and describe how our contributions add to this work. In x3, we discuss the theoretical foundations motivating the approach, and in x4 we describe the methodology in detail. We describe our validation methodology in x5 and the main results in x6. x7 provides a series of sensitivity analyses. Finally in x8, we summarize the implications of this work, note its limitations, and provide recommendations for future research. 2. Background and Related Work 2.1. Brand Attribute Ratings Marketing managers have long relied on estimates of consumer perceptions of brands along attributes of interest to inform marketing strategy (John et al. 2006, Lancaster 1971, Lehmann et al. 2008). Perhaps most notably, such estimates are used as the primary input for generating perceptual maps, which have been used by managers since at least the 1970s to understand the relative positioning of competitive brands (Hauser and Koppelman 1979, Johnson and Hudson 1996), and are widely considered a foundational analytical tool in marketing research (Green et al. 1989, Shocker and Srinivasan 1979, Steenkamp et al. 1994). Developing improvements to perceptual mapping techniques consistently remains a priority for marketing researchers (Bijmolt and van de Velden 2012, Day et al. 1979, Dillon et al. 1985, Kaul and Rao 1995). Researchers have proposed both compositional and decompositional techniques for elic- iting brand ratings from consumers. Both require recruitment and interaction with a large Culotta and Cutler: Mining Brand Perceptions from Twitter Social Networks 4 Article submitted to Marketing Science; manuscript no. (Please, provide the manuscript number!) and diverse set of participants; in the former, users are asked to directly rate brands on a numeric scale according to their strength in a given attribute, while in the latter, users are asked to perform sorting tasks on brands, from which attribute ratings are inferred (Huber and Holbrook 1979). Some research suggests that compositional methods (i.e., rat- ing brands via surveys) can provide greater validity (Bottomley et al. 2000, Hauser and Koppelman 1979), and these are widely used in marketing practice (Steenkamp et al. 1994). Yet, compared to the wealth of research focused on the advancement of techniques for making inferences from such brand attribute ratings, surprisingly little research has focused on advancing methods for obtaining the ratings themselves (Steenkamp and Van Trijp 1997). At the same time, many researchers have called out substantial limitations result- ing from the requirement of collecting primary data to inform these analyses, including difficulty and expense in recruiting sufficient participants, and in maintaining participants attention and cooperation during tasks (Day 1975, McDaniel et al. 1985, Steenkamp and Van